Modeling What Persists: Learning Explicit Affective States for Character Simulation

Simulating a character across extended interactions requires more than producing locally appropriate responses; some aspects of the character must persist over time. Existing systems largely condition generation on profiles, memories, and dialogue history, while affect is often inferred only for the current response. This raises a basic question: should affect be treated as a transient generation cue or as a persistent state of the simulated character? We study the latter and propose Unified Explicit Interoception Modeling (UEIM), which separates affective state tracking from behavior realization. UEIM maintains a continuous Valence–Arousal state anchored by fine-grained emotion prototypes and updates it with each interaction event. Because such latent states lack reliable continuous supervision, we introduce behavior-grounded state–behavior alignment: teacher-derived trajectories initialize state transitions, while a frozen behavior generator and affective reader provide proxy feedback through observable affective consequences. Experiments on DailyDialog and ESConv show gains in response realization and affective transfer; frozen-A1 studies further improve negotiation success and affective attunement in counseling and customer service. These results support modeling not only what emotion should be expressed now, but also what affective state has persisted and how the current event changes it.